Recognition and Segmentation of the Earth's Surface Using Convolutional Neural Networks

The work examines the influence of various factors on the result of training convolutional neural networks when solving the problem of satellite image segmentation. Different approaches to solving the problem were considered, including the use of different configurations of convolutional neural networks and their different architectures. It has been established that using a neural network of the U-Net architecture with a ResNet-34 encoder, using a combination of CCE and a soft Jaccard index in equal proportions as a loss function, shows the best results for the task under consideration; it has also been established that using the LeakyReLU activation function gives the best result for the Stacked U-Nets architecture and that using a Gaussian filter to process input images do not have a positive effect.